dsh-computer-use: A Computer Use Capability Plugin for DeepSeek Harness

What Does It Do?

dsh-computer-use is a DSH plugin for DeepSeek Harness that gives models the ability to operate a computer through an isolated browser or native Windows desktop. It solves the problem of models being unable to directly control a browser or desktop app by providing a unified capability seam and six model-facing tools, enabling observation, clicking, typing, and takeover. The plugin supports multiple providers (fake, playwright, windows) and integrates third-party vision models for perception, while enforcing risk-classified approval and domain allowlisting for safe operation.

Core Features

• Unified computerUse capability: start, listTargets, observe, act, stop operations exposed to models.
• Multiple backends: fake (tests), playwright (isolated Chromium), windows (native desktop helper, Windows-only).
• Dual perception modes: accessibility (no image) and analyze (screenshot → third-party vision model → structured result).
• Safety mechanisms: risk classification, fail-closed approval, domain allowlist, and full replayable session log.
• Chrome Cookie Bridge: import logged-in browser sessions to reuse existing auth.
• Live configuration: toggle the capability, switch provider, and adjust browser options directly from the DSH GUI without editing YAML.

How to Use This Plugin?

After enabling it in DSH, models can control the browser or desktop via tools like computer_observe and computer_act. Configuration lives on the dsh-computer-use/plugin row (config:), for example in cordis.patch.yml set provider to playwright, visionProvider to xiaomi, visionModel to mimo-v2.5. Alternatively, use the DSH Settings → computer-use section to adjust enabled, provider, and browser options in real time. For Windows desktop operations, use the windows provider and configure the native helper command.